Re: [PATCH bpf-next v2 0/7] Add support for XDP_ATTACH

From: "Björn Töpel" <bjorn.topel@intel.com>
To: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: "Magnus Karlsson" <magnus.karlsson@gmail.com>,
	"Jesper Dangaard Brouer" <brouer@redhat.com>,
	"Björn Töpel" <bjorn.topel@gmail.com>,
	"Karlsson, Magnus" <magnus.karlsson@intel.com>,
	ast@kernel.org, "Daniel Borkmann" <daniel@iogearbox.net>,
	"Network Development" <netdev@vger.kernel.org>,
	"William Tu" <u9012063@gmail.com>,
	"Zhang, Qi Z" <qi.z.zhang@intel.com>,
	"Jakub Kicinski" <jakub.kicinski@netronome.com>,
	andrew@lunn.ch
Subject: Re: [PATCH bpf-next v2 0/7] Add support for XDP_ATTACH
Date: Wed, 19 Dec 2018 10:34:28 +0100	[thread overview]
Message-ID: <a0e6c2cb-f2c8-ba7c-7c72-c20826b84172@intel.com> (raw)
In-Reply-To: <20181218230358.4bs53pwg65f2kdtc@ast-mbp>

On 2018-12-19 00:04, Alexei Starovoitov wrote:
> On Mon, Dec 17, 2018 at 04:30:05PM +0100, Björn Töpel wrote:
>>>
>>> Instead of introducing the XDP_ATTACH option to the bind call, can we
>>> just make this association between rx queue id and xsk every single
>>> time we bind? Then it is up to the user via the XDP program if he/she
>>> wants to use this by calling xsk_redirect. No new option needed.
>>>
>>
>> Nice! Then it would simply be a matter of adding the helper. Much better
>> than extending the uapi. Thank you for pointing this out!
> 
> Magnus's by-default association approach sounds good to me.
> I'm missing though why extra bpf helper is needed.
> This soft-bind rx queue with xsk at the time
> of normal bind() should work as-is and no program necessary.

Hmm, an XDP program has always been required. It's just whether the
program needs an XSKMAP or not. The XDP_ATTACH can be seens as "use an
internal map" -- and the mechanism to redirect using that map is via the
bpf_xsk_redirect helper. What Magnus proposed was populating the builtin
map unconditionally, and therefore removing the need for the XDP_ATTACH
option.

If I am interpreting what you're saying correctly, you thought
XDP_ATTACH (in v1) was bypassing the XDP path. This is not something
that we want. AF_XDP is a proper XDP construct, and require an XDP
program to receive frames. What I wanted to improve with this series was
to lower the complexity of the minimal XDP program to receive a frame.

> No prog -> default assoc. Yet the prog has an ability to redirect
> to specific xsk via existing bpf_redirect_map
> or via faster (future) xsk_redirect.
> 
> Also I agree with Jesper that per-queue programs would be good to have
> sooner than later, but I disagree that it's mandatory for this set.
> It seems the patches will give us performance boost and simplification
> of uapi while per-queue prog api is being discussed and implemented.
> 
> I like the other idea too (of having libbpf provide canned bpf progs
> for different use cases). That's better than loading them
> from the kernel.
>

To summarize; I'll hold off this series for now (except the first
cleanup patch which I'll send out today), and spend some time on
per-queue programs, iproute2 support and libbpf.

Björn